Deploying the Apache HTTP Server Within the Apache Software Foundation
Total Page:16
File Type:pdf, Size:1020Kb
Deploying the Apache HTTP Server within the Apache Software Foundation Justin R. Erenkrantz University of California, Irvine http://www.erenkrantz.com/oscon/ [email protected] Why should I pay attention? One of the folks behind [email protected] Responsible for maintaining servers Committer to Apache HTTP Server and APR Familiar with the httpd 2.x codebase Committer to Subversion (we run that too) ‘Spare’ time: PhD student at UC Irvine What is the ASF? The Apache Software Foundation (ASF) Organizational, legal, and financial support Not-for-profit foundation Currently has 24 top-level projects (TLP) Each TLP may have many codebases All of these have websites, code, mailing lists Apache HTTP Server History Apache HTTP Server has been market leader for over eight years...and counting 2.0 series went GA in April 2002 New architecture to fix real-world issues Windows (and others) are first-class now Includes mod_dav, mod_ssl out of box What does the ASF provide? Three essential services: Websites: http://jakarta.apache.org/ Version control: httpd-2.0 repository Mailing Lists: [email protected] The failure of any impairs the project! What does the ASF use? Desire to eat own dogfood, if possible Websites: Apache HTTP Server 2.0 Version control: CVS and Subversion Mailing lists: ezmlm, qpsmtpd, qmail Platform: FreeBSD (4.x and 5.x) What level of service? Users: ~900 committers & ~130 members Websites: ~40 million views/month Version control: ~8GB of source Mailing lists: ~400k inbound emails a day Expect these numbers to continue to rise Organizational Structure Email is key to our coordination planning IRC used for real-time fire-fighting infrastructure@: General list root@: People with root access apmail@: People who can create lists Physical access: Big red button hitters Where do the servers reside? Like our contributors: all over the place Main servers now at UnitedLayer in SF Used to reside at CollabNet (for free!) Strong desire to be self-reliant Pay our own co-lo costs now Can now provide access to those in area Supporting Hardware Still try to reduce need for physical access Terminal Server Allows remote serial console Not all machines support BIOS output! Power distributor Allows power-cycling if all else fails The ASF machines minotaur: Shell, CVS, WWW hermes: email server nagoya: bug-tracking, web mail archives brutus: Gump continuous builds moof: Apple test box ajax: European backup The ‘Not-quite-so-big’ Picture minotaur (CVS, WWW) Cyclades ACS nagoya: E4500 (Webapps) hermes: x345 (SMTP) eris: x345 (Webapps?) brutus: x345 (GUMP) loki: x345 (VMWare?) Cyclades PM20 N X moof: PowerMac (Testing) ajax: Itanium (Backup) Dual Virtual Host Strategy www.apache.org = cvs.apache.org Traffic primarily on www instance ‘cvs’ requires Subversion, SSL, WebDAV High(er) memory footprint! Leverage ‘IP alias’: .194 and .195 Two instances with one optimized for space Two Parallel Instances Other advantage of two instances Allows ‘unstable’ testing, too www.apache.org httpd-2.0 APACHE_2_0_BRANCH cvs.apache.org httpd-2.0 HEAD (aka 2.1) Choice of MPM Multi-Processing Module (MPM) in 2.0 Prefork: Traditional Apache 1.3 model Worker MPM: smaller footprint w/ threads Configured for 800 clients. Yet, we use prefork...Why? Threads busted in FreeBSD (5.3: first fix!) Off-loading Large Transfers Main webpages are not mirrored Downloads are mirrored Selected Mirroring Strategy Importance of being fresh: 4x daily. ~5GB of releases served by mirrors ~225 mirrors around the world The Need for Mirroring ASF had long history of informal mirrors Only HTTP Server used them Jakarta was majority of traffic Users defaulted to our servers not mirrors Too many downloads to serve ourselves Had to respect our bandwidth cap! Mirroring Strategy Made mirrors mandatory Looked at how SourceForge did it Found extra steps too cumbersome... Promote digital signatures Mirrors could be corrupt or malicious Signature links still point to our servers Building a Web of Trust MD5 and PGP signatures available Provide that it hasn’t been corrupted Yet also requires user participation You need to verify the signature You should also have a path to signer Keysigning parties are extremely beneficial Mirroring Overview Developed Python script to read mirrors Greg Stein’s EZT template engine Project-specific look and feel Use GeoIP to find geographic region Present locale-appropriate mirrors 95% percentile usage now ~18Mbps Evolution of ASF Setup Then: daedalus (mail/www); icarus (CVS) Now: minotaur (www/CVS); hermes (mail) Icarus retired when minotaur came Begin of switchover to UL co-lo Hermes forced into production Had been testing it, but daedalus died The role of minotaur daedalus and icarus were initial boxes Dual PIII/800s: ~2000-2003/2004 Obtained around early 2003 Xeon/2.4GHz with HyperThreading RAID5 array with 400GB Two months doing ‘make world’ Why was that history important? We do not have a pure dedicated web server Most compete with other critical services In minotaur’s case, CVS is hosted on it Also acts as shell server for accounts Our CVS setup mandates shell accounts Must keep load ‘low’: ~1 load average (CVS) No dynamic pages Emphasize static content to reduce load Adopt tools like Anakia, Forrest, etc. Transform XML into (X)HTML Provides benefits of SSIs without cost! Python CGI scripts handle mirroring MoinMoin Python Wiki (wiki.apache.org) Deciding When to Deploy Two reasons for deploying a new build New release pending Prefer to see a release run for 48 hours Minotaur has ‘honorary’ release vote Resolve issues seen on our servers Fixes for unknown reproduction cases Deploying New Builds Greg Ames and Jeff Trawick handle www Will deploy a new build and send email Only one custom patch at the moment Stores input in brigade Facilitates crashdump reproduction Rest of patches have been committed! Responding to Failures httpd architecture is fairly resilient Each client handled in separate process Crashes cause new child to be spawned Segfault triggers crash dump and log Most common install error with ‘suexec’ We often forget the suid root bit! (Oops) Adopting Subversion Moving away from CVS to Subversion Developer advantages: Renames, atomic commits, etc, etc. Administration advantages: Easier incremental backups Better access control Subversion Backups CVS has fundamental flaw Every commit changes an RCS file Not possible to keep just ‘delta’ SVN: Incremental backups post-commit Atomic commit can be ‘replayed’ later Synchronized to off-site mirrors Access Control As ASF has grown, blurred group lines Avalon wants to give commit access to Cocoon and JAMES developers Complicated ‘avail’ system on top of CVS mod_authz_svn provides group control First written by Sander Striker (root@ too) Migrating from CVS More projects are starting to ask to migrate No forced migration...yet. Users beginning to feel comfortable Tried to make all tools available from CVS ViewCVS still works, commit emails, etc. Still haven’t obtained a valid SSL cert! Ugh. Spam, spam, spam Spam has been an increasing problem Most email is for mailing lists Human moderator gets requests Hundred moderate emails/day for some Reach tipping point for us Re-deployed a new mail architecture qmail I wish I knew why we use qmail...Yet, we do. Have lots of hacks to handle our load Remote Concurrency & Big ToDo Very hard for new admins to understand Can’t migrate away from qmail easily ezmlm is too central to our mail delivery qpsmtpd Replacement for qmail SMTP component Written in Perl by Ask Bjørn Hansen perl.org and mysql.org uses it too Allows easy introduction of plugins Major difficulty with qmail by itself If you use qmail, highly recommended! clamav Free GPL Virus Scanner Automatic updates through freshclam Daemon via persistent clamd Qmail-scanner was awful with clamav Actually spurred us to qpsmtpd Rejects about 10,000 messages a day SpamAssassin Part of our strategy to eat our dog food Using SA 3.0.0rc1 with spamd Reject if email over 10.0 score All network tests enabled now No Bayes rejection...yet (How??) Rejects about 12,000 messages a day Real-time Blackhole Lists Reject a message outright if you are on: Spamhaus XBL/SBL SORBS DUL RBL (Dynamic IP) DSBL Rejects about 130,000 messages a day! Wish didn’t have to reject dynamic IPs, but... Custom Plugins check_virtualdomains: Mail to @apache.org check_badrcpto: Reject john@, clark@ exe_filter: Blocks EXE & ZIP (~90k/day!) spamwatch: Custom rule sets SpamAssassin guys showed not effective Catches ~500 emails/day w/false positives Forced migration! We received a donation of four IBM 345s Did initial testing; No urgent schedule Daedalus hard drive died... Placed hermes into production before ‘ready’ Later found problem with RAID controller Deployed a kernel patch on faith (worked) Lessons Learned We can do 40 million page views/month on one box. You can too. Be smart! Always re-evaluate what you are doing. Try to involve as many as feasible. Can you off-load the work? Stick to the basics. Useful links WWW: http://www.apache.org/server-status/ CVS: http://cvs.apache.org/server-status/ Henk Penning: http://www.apache.org/~henkp/ Vadim Gritsenko: http://www.apache.org/~vgritsenko/.